A brain-inspired computer can now simulate part of the sensory cortex in real time, using tens of thousands of virtual neurons. It is the first time such a complex simulation has run this fast and could be an important step towards building聽better brains for robots.
The SpiNNaker supercomputer at the University of Manchester, UK, features 57,000 specialised chips with a total of 1 million processing units, known as cores. It is designed to run programs that simulate how biological neurons behave and interact, which could help neuroscientists experiment on virtual brain circuits to study disease and cognition.
The computer shuttles information around in a similar way to the brain, says Oliver Rhodes, who led the research. Standard supercomputers send big blocks of data at set times, but SpiNNaker鈥檚 cores can transmit small blocks to hundreds of other cores simultaneously whenever required.
Advertisement
Now Rhodes鈥檚 team has shown that SpiNNaker can simulate the behaviour of 77,000 neurons in the sensory cortex, equivalent to about 1 square millimetre鈥檚 worth, as fast as it happens in the brain for up to 12 hours. 鈥淲e can run this at the same speed as biology, which is quite a big achievement,鈥� he says.
Biological neurons process incoming signals from many other neurons to decide whether to send signals themselves, and as networks get bigger, the input to each neuron skyrockets. That bogs down supercomputer simulations where each processor typically deals both with incoming signals and with deciding if neurons should fire.
So Rhodes split the work up, dedicating some SpiNNaker cores to computing neuron behaviour but even more to processing inputs. That let the simulation run faster and will enable bigger simulations to run in real time, he says.
This particular program lets people model how networks of neurons interact rather than computing information, says Rhodes, but it can be a building block for more complex models of things like sensory processing.
The research is a big step towards large-scale simulations of brain processes, says Brad Aimone at Sandia National Laboratories in the US. Researchers will have to run hundreds of simulations to confirm results, so making them fast is crucial. 鈥淣ow the burden is back on to the neuroscientists to actually build models that can scale to this level,鈥� he says.
Markus Diesmann at the J眉lich聽Research Centre in Germany, who designed the model used by SpiNNaker, says real-time operation means the chips could also be useful in robotics. 鈥淵ou can really transfer principles and algorithms that you鈥檝e uncovered in nature into this artificial brain living in a robot,鈥� he says.
Diesmann鈥檚 group has now built a model of the visual system made up of millions of neurons and is working with the Manchester team to port it over to SpiNNaker.
搁别蹿别谤别苍肠别:听arXiv,听
